Monitoring, diagnostic and prognostic methods - VII
Chair: Dr. Yan-Hui LIN (Beihang University)
Probabilistic Forecasting Method for Concrete Creep Prognosis under Model-form Uncertainty Dr. Seung-Seop JIN, Sang-Lyul CHA and Hyung-Jo JUNG (KAIST)
Creep and shrinkage are the time-dependent deformation of structural concrete under the influence of sustained strains generated by external conditions. Time-dependent deformation of structural concrete may cause cracks and has much to do with durability performance of structural concrete. Therefore, it is important to make a reliable prognosis for time-dependent deformation. For predicting creep and shrinkage, several empirical models (competing models) are available by different professional societies: ACI, fib 2010, KCI and so on. Consequently, model-form uncertainty inevitably arises from the competing models. In addition, the differences between the predictions of competing models are significant. To consider the model-form uncertainty, we propose a new method to make a reliable prediction for the creep and shrinkage with the experimental data. The proposed method uses trans-dimensional Markov Chain Monte Carlo algorithm to quantify the model-form uncertainty and parameter uncertainty simultaneously. Based on quantified model-form uncertainty, creep and shrinkage behaviors are predicted by BMA. We present the predictive performance of the proposed method by experimental data of creep and shrinkage tests.
Running State Evaluation Method of Turbine Bearing Based on Feature Vector Dr. Hao ZHANG, Pengfei DANG and Qingkai HAN (Dalian University of Technology)
Due to the high modal coupling of the turbine bearing in environment control system, it is very difficult to extract of the vibration signal feature and construct the recognition model on different feature. A running state evaluation method of the turbine bearing is proposed based on the feature vector with limited testing data in this paper. Firstly, aiming at some failure modes in several typical faults of turbine bearing, three time domain feature parameters and seven frequency domain feature parameters are chosen to construct feature vector for discrimination. Then, the feature vectors of different fault testing data are dimensional reduced based on the principal component analysis method. Based on above, the support vector machine(SVM) model of the turbine bearing running state, as well as the auto-regressive and moving average (ARMA) model of the turbine bearing failure prediction are proposed for monitoring and predicting the occurrence and development of typical turbine bearing failure modes. Experimental results suggest that the bearing running state evaluation method proposed in this paper can improve the prediction accuracy effectively.
Development of Easy Maintenance Assistance Solution (EMAS) for Gas Turbine Dr. Jayoung KI and Myoungcheol KANG (EGT Co., Ltd)
The solution was developed for the maintenance decision support of combined cycle power plant gas turbine. The developed solution was applied to MHI501G gas turbine and is, in present, on the process of field test at GUNSAN combined cycle power plant, South Korea. The developed solution provides the calculated result of optimal overhaul maintenance period through following modules: Real Time Performance Monitoring, Model-Based Diagnostics, Performance Trend Analysis, Optimal Overhaul Maintenance Interval, Compressor Washing Period Management, and Blade Path Temperature Analysis. Model-Based Diagnostics module analyzed the differences between the data of gas turbine performance model and the online measurement. Compressor washing management module suggests the optimal point of balancing between the compressor performance and the maintenance cost.
Deep Stack Dictionary learning for Fault Diagnosis of Rotating Machinery Dr. Jinyang JIAO, Ming ZHAO and Jing LIN (Xi'an Jiaotong University)
Effective fault diagnosis of rotating machinery has been a hot topic in the prognosis and health management. However, it is a challenging problem to extract periodic impulses under heavy background noise. In the last decade, deep learning and dictionary learning have been promising methods to extract feature information, which have made great achievement in the field of image, video denoising, etc. In this paper, via fusing the deep learning with dictionary learning, an algorithm called deep stack dictionary learning is proposed. This algorithm is trained in a layer-wise greedy fashion so as to get rid of the noise and obtain better periodic impulses. The proposed algorithm mainly has two merits: 1). Dictionary learning can adaptively learn fault feature from the original signal without any prior knowledge. 2). The deep stack strategy further improves the feature of learned periodic impulses, and extract weak fault feature in the early stage successfully. The effectiveness of the proposed method is validated through numerical simulation and datasets from rolling element bearings. Compared with single layer dictionary learning, the deep stack dictionary learning can extract periodic impulses and eliminate background noise effectively.
Parameter Estimation Using Particle Filter for Induction Machines under Inter-Turn Fault Mr. Viet Hung NGUYEN, Danwei WANG, Jeevanand SESHADRINATH, Abhisek UKIL, Viswanathan VAIYAPURI and Sivakumar NADARAJAN (Nanyang Technological University, Rolls-Royce Singapore Pte. Ltd.)
Parameter estimation has found its applications in various domains. In this paper, it is applied to fault severity estimation. A method, using particle filter approach, for estimating unknown fault parameters in stator winding inter-turn short, is firstly proposed. These parameters are insulation resistance and percentage of shorted turns. The method uses only measurements of stator voltages and currents. In order to effectively estimate the parameters, a multiple-model approach is exploited. A sequence components-based approach is applied to derive an equality constraint on the magnitude of a state variable, which works as an additional information for estimation algorithm based on state-state model. Additionally, the variance reduction technique is applied to increase the accuracy of the method.